Pi coding agent integration for Helix via Steel plugins.
✅ Working - Streaming responses, session persistence, prompt caching
[pi/input], [pi/output]) — nothing else.
Use pi CLI directly to change models.
Custom Helix build: mattwparas/helix (Steel fork) with PR #8546 (window resize/focus mode).
cd ~/git/helix && cargo install --path helix-term --lockedmkdir -p ~/.config/helix/cogs/pi
cp src/*.scm ~/.config/helix/cogs/pi/
# Configure helix.scm (see docs/tutorial.md):pi-start # Start session
:pi-send # Send prompt
:pi-quit # Close session
Almost all of this code is AI/LLM-generated. It's best used as a source of inspiration for your own AI/LLM efforts rather than as a traditional library.
This is personal alpha software. All my GitHub projects should be considered experimental. If you want to use them:
- Pin to a specific commit — don't track
main, it changes without warning - Use AI/LLM to adapt — without AI assistance, these projects are hard to use
- Treat as inspiration — build your own version rather than depending on mine
Suggestions welcome — If you have ideas for improvements or changes, I'd be delighted to read them and use them as inspiration for my own efforts.
Why not a library? These days it's often quicker to use AI/LLM to build your own than to integrate traditional libraries. My use of AI/LLM is inspired by these people and posts:
- Simon Willison's Weblog — Essential reading on LLMs, prompt engineering, and building with AI
- CLI over MCP — Armin Ronacher on why command-line tools are better integration points than custom protocols
- Build It Yourself — Armin Ronacher: "With our newfound power from agentic coding tools, you can build much of this yourself..."
- Shipping at Inference Speed — Peter Steinberger on the new workflow of building with AI assistance
- Year in Review 2025 — Mario Zechner on AI-assisted development
What I use: Currently Anthropic's Claude Opus, evaluating OpenAI's GPT Codex as an alternative.
Dual-licensed under MIT and Apache-2.0.